Extract Table Pagination

When to invoke

When the user task involves reading structured data from an HTML table — the page either shows a <table> element, or a role="table" ARIA structure. If the table is paginated, the skill walks pages via a "next" link/button until either the next link is absent, the row count stops growing, or a configurable max_pages is reached.

The matcher gives this a high score when the DOM summary contains <table and the visible-text sample has table-like indicators (numeric columns, repeated short text).

Recipe

1. wait_for_dom_ready timeout=5s
2. wait_for_selector selector="table, [role='table']" state=visible timeout=5s
3. extract_table selector="table" into="$rows_page_1"
4. wait extra=200ms

Success criteria

• assert $rows_page_1 is_non_empty_list

Pagination companion

For paginated tables, chain with the pagination-next-page skill in a loop. The runner's vars map preserves bindings across skill invocations so the agent can accumulate rows.

Pseudo-recipe the agent assembles:

while page <= max_pages:
    invoke_skill("extract-table-pagination") → $rows_page_N
    accumulated += $rows_page_N
    if no "next" link visible: break
    invoke_skill("pagination-next-page")

We deliberately keep extraction and pagination as separate skills so the agent can interleave with other steps (rate-limit pauses, filter applications, deduplication).

When NOT to use

• The data lives in a <div>-grid layout (card grid, masonry, etc.) — use the future extract-card-grid skill, or vision fallback.
• The table is virtualized (only renders visible rows on scroll, e.g., AG Grid, MUI DataGrid). Use handle-infinite-scroll with row-extraction inside the scroll loop instead.
• The "table" is actually a layout <table> from 2003-era HTML (no <th>, no semantic structure). The extractor will produce noisy results; better to extract by text-selector and structure manually.

Known failures

• Virtualized tables (~5% of modern web apps): only the visible rows are in the DOM at any time. The extractor returns just those. Workaround: combine with handle-infinite-scroll or scroll-and-extract.
• Tables with rowspan/colspan (~3%): the naive header-zip extractor mis-aligns columns. Vision fallback handles for now; v0.2.0 will support extract_table mode="rowspan-aware".
• No <thead>, no <th> (~5% of legacy sites): the extractor uses the first row as headers; if that row is data, the first record is lost. Workaround: pass headers=[...] arg (planned v0.2.0).
• Server-side-rendered "infinite scroll" disguised as pagination: the next-page link returns the SAME rows on click. Detection logic in pagination-next-page aborts when row hashes match.

Example usage

Agent task: "Get the latest cs.AI papers from arxiv."
→ invoke_skill("verify-page-loaded") → success
→ invoke_skill("extract-table-pagination") → success, $rows_page_1 = [{title, authors, abstract}, ...]
→ agent inspects $rows_page_1 and reports

Related skills

• verify-page-loaded — run first; tables on slow-rendering pages need readiness check
• pagination-next-page — chain to walk multiple pages
• handle-infinite-scroll — alternative when content lazy-loads instead of paginating
• dismiss-cookie-banner — usually run before any extraction